This repository is associated with the FoREST: Frame of Reference Evaluation in Spatial Reasoning Tasks. paper. This repository provides the code for context generation and the evaluation tasks related to FoREST.
Please run the below command to create the environment,
# Create a conda environment
conda create -n FoREST python=3.11
conda activate FoREST
# Install requirements
pip install -r requirements.txt
Pipeline for dataset creation, starting from selecting a locatum and relatum from available objects and then applying a spatial template to generate the spatial expression (T).
FoRs are assigned based on the relatum’s properties. Then, based on the number of FoRs, T is categorized.
For example, A cat is to the right of a dog. (with two possible FoRs: external intrinsic and external relative) belongs to the A-split.
Then, its disambiguated version (A cat is to the right of a dog from the dog's perspective) is added to the C-split. Next, if applicable, a relatum's orientation is included for visualization and question generation.
Finally, Unity3D generates scene configurations, creating question-answer pairs from T.
We provide the relatum properties used for creating each Frame of Reference case below (Note that the images are only for visualization),
The visualization is created separately using the Unity3D engine using free models downloaded from assert store. The developed code is provided in this repository.
Examples of synthetic images are provided below,
All images will be downloadable using the hugging fact dataset (This will be updated soon).
We utilize the FoREST benchmark to evaluate three tasks. The two main tasks focus on spatial reasoning: question-answering (QA) and text-to-image (T2I). All evaluation tasks can be found below,
If you find this repository useful, please cite
@misc{premsri2025forestframereferenceevaluation,
title={FoREST: Frame of Reference Evaluation in Spatial Reasoning Tasks},
author={Tanawan Premsri and Parisa Kordjamshidi},
year={2025},
eprint={2502.17775},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2502.17775},
}